As the world transitions rapidly toward clean electric mobility, electric two-wheelers and light EVs have become the backbone of urban transit. However, most modern light EV instrumentation remains fundamentally reactive—displaying simple battery percentage bars and generic warning lights only after a critical fault or breakdown has already occurred. Furthermore, in severe collision scenarios where the main high-voltage traction rail is severed or tripped by safety contactors, the vehicle loses all power, leaving riders stranded without emergency communication.

EVA Guardian is an intelligent, distributed edge-computing platform designed to make electric vehicles truly self-aware, proactive, and resilient. Built for the element14 EZ-EV Design Challenge,

EVA Guardian Features:

Distributed Dual-Brain Architecture: Decouples battery analytics from incident monitoring across two dedicated Arduino UNO Q microcontrollers communicating over an isolated differential RS-485 bus.
On-Device Edge AI Motion Classifier: Trained with Edge Impulse to distinguish normal riding dynamics (turns, acceleration, speed breakers, potholes) from actual crashes and tip-overs with sub-5ms latency and zero cloud dependence.
Hybrid State of Charge (SoC) Engine: Combines real-time Coulomb counting with automatic 60-second Open-Circuit Voltage (OCV) relaxation recalibration, eliminating voltage sag distortion and sensor drift.
Automated 4-Step Hardware Fault Diagnostics: An intelligent protocol that autonomously isolates whether a communication blackout is caused by the master transceiver, slave transceiver, or a severed harness cable.
Predictive Connector Degradation & Auto-Failover: Real-time dual-point voltage sensing detects contact resistance buildup (ΔV > 1000mV) before thermal damage occurs, coupled with a seamless redundant backup battery path governed by the core rule: "If transportation stops, communication must not."

Figure 1 — EVA Guardian Complete Hardware Setup: Battery Management Node, Incident Detection Node, RS-485 Differential Bus, and Onboard Visual Displays.

EVA GUARDIAN COMPLETE SYSTEM ECOSYSTEM

Node 1: Battery Management & Profiling	Node 2: Incident Detection & Safety
• 2S Li-ion Primary Pack (7.4V - 8.4V)	• 2S Li-ion Backup Pack (Independent Path)
• HX-2S-A2 Hardware Protection Board	• Automatic Power Transfer Relay Circuit
• LF412 High-Impedance Op-Amp Buffer	• GY-91 6-DOF IMU (MPU6500 @ 42 Hz)
• ACS723 Hall-Effect Current Sensor (±5A)	• Edge Impulse On-Device Neural Network
• 10k NTC Thermistor (Steinhart-Hart)	• 3-Window Temporal Accident Verification
• Hybrid SoC (Coulomb + 60s Rest OCV Sync)	• Matrix LED Spirit Level & Flashing SOS
• Zephyr RTOS Mutex-Protected Matrix Display	• Real-time Telegram SOS Bot & WebUI
• ADM3068E RS-485 Slave Node	• ADM3068E RS-485 Master Controller
Differential Pair RS-485 Bus [ Autonomous 4-Step Fault Diagnostic Routine ] • [ Real-Time Connector Voltage Differential ΔV ]

The Core Problem: Beyond Static Battery Bars

Most modern electric two-wheelers display two main numbers: Battery Percentage (%) and Estimated Range (km).

Under real riding conditions, however:

1. Dynamic Voltage Sag Misleading the Rider

When accelerating or climbing hills, cell internal resistance causes terminal voltage to drop significantly. Simple voltage-based meters report a false battery drop, confusing riders.

2. Invisible Harness & Connector Aging

Over months of road vibrations and environmental exposure, crimps and high-current connectors oxidize. Contact resistance increases, leading to I²R localized power dissipation, terminal melting, and fire risks that standard BMS boards never flag.

3. Brittle Crash Detection

Simple acceleration threshold triggers produce massive false-alarm rates over bumpy terrain and potholes.

4. Post-Crash Communication Blackout

When a crash occurs and the main battery breaker trips, vehicle telematics die instantly, leaving injured riders unable to broadcast emergency location and crash telemetry.

EVA Guardian was engineered from the ground up to solve each of these failure modes.

System Architecture & Design Philosophy

Rather than routing all sensors to a single centralized microcontroller—creating a single point of failure—EVA Guardian decouples vehicle intelligence into two modular nodes:

Figure 2: EVA Guardian High-Level Technical Architecture

Figure 3 — Detailed Technical Architecture: Signal conditioning front-ends, dual Arduino UNO Q SBCs, RS-485 bus, and redundant battery failover routing

The Three Architectural Pillars:

Decoupled Responsibilities: Battery chemistry tracking and vehicle motion dynamics operate at fundamentally different timescales (10 Hz vs. 42 Hz) and are managed by independent processors.
Differential Industrial Bus (RS-485): High-current motor controllers generate severe electromagnetic interference (EMI). Differential signaling over twisted pair via Analog Devices ADM3068E transceivers ensures zero packet loss in harsh electrical environments.
Isolated Redundant Power Architecture: The Safety Node contains its own power management circuit and backup 2S battery pack, completely independent of the main traction bus.

Hardware Engineering & Analog Signal Conditioning

Precision analytics require clean, calibrated sensor signals. The analog front-end (AFE) of the Battery Node was designed to maximize signal integrity on the Arduino UNO Q's 14-bit ADC (0−16383 counts).

Figure 4: Precision Measurement Circuitry on Veroboard: LF412 buffer stage, ACS723 Hall-effect sensor, NTC bridge, and HX-2S-A2 protection module.

1. High-Impedance Pack Voltage Buffer (LF412)

Directly connecting a resistive divider to an ADC causes impedance loading errors. A precision divider (R1 = 6.8 kΩ, R2 = 3.0 kΩ) scales the 8.4V peak voltage down to a safe 2.57V. An LF412 dual JFET op-amp configured as a unity-gain voltage follower (10¹²Ω input impedance) isolates the divider, preventing ADC sampling charge-injection from skewing the measurement:

Vbattery = VA1 × (6.8 kΩ + 3.0 kΩ) / 3.0 kΩ = VA1 × 3.2667

2. Isolated Bidirectional Current Measurement (ACS723)

An Allegro ACS723 Hall-effect linear current sensor is placed in series with the battery load:

Galvanic isolation eliminates ground loop noise.
Calibrated zero-current quiescent voltage: V_ref ≈ 1.70 V.
Sensitivity: S ≈ 1.336 V/A.

Ibattery = (VA0 − 1.70 V) / 1.336 V/A

(Negative current indicates discharge through load; positive indicates charging).

3. Thermal Acquisition via Steinhart-Hart Equation

An NTC thermistor (10 kΩ nominal @ 25°C) is thermally bonded to the cells. The firmware converts resistance R_NTC to temperature using the Steinhart-Hart β-parameter model:

1/T = 1/T0 + (1/β) × ln(RNTC / R0) ⇒ TCelsius = T − 273.15

Edge AI Motion Classification & Incident Detection

To eliminate false crash alarms caused by speed bumps and potholes, the Incident Detection Node utilizes on-device Edge Machine Learning.

Figure 5 — Edge Impulse Studio: 39-feature Power Spectral Density (PSD) extraction, 5-class motion dataset distribution, and validation confusion matrix.

1. Dataset Collection & Labeling

Using the GY-91 (MPU6500 6-axis IMU) sampled at 42 Hz, approximately 11 minutes of real-world motion data was recorded across 5 distinct operational classes:

Accident: High-g impact followed by rapid roll/tip-over and sustained tilt.
Idle: Stationary with ambient engine/road rumble.
Front and Back: Longitudinal acceleration and regenerative braking.
Right and Left: Slalom cornering and banking angles.
Up and Down: Transient vertical shocks from speed breakers and potholes.

2. Spectral Analysis (Power Spectral Density - PSD)

Before feeding the neural network, the raw IMU stream passes through an Exponential Moving Average (EMA) filter (α = 0.25). An Edge Impulse Spectral Analysis block extracts 39 frequency-domain features, capturing the unique harmonic fingerprints of impacts versus bumps.

3. Neural Network Architecture

The classifier runs fully on-device on the Arduino UNO Q without cloud connectivity:

Input Layer: 39 extracted PSD spectral features
Dense Layer 1: 20 neurons (ReLU)
Dense Layer 2: 10 neurons (ReLU)
Dense Layer 3: 5 neurons (ReLU)
Output Layer: 5 neurons with Softmax probability distribution
Training Parameters: 50 epochs, learning rate = 0.0005 Adam optimizer.

4. Interactive Electronic Spirit Level on LED Matrix: During normal operation, the Arduino UNO Q's 8×12 LED matrix dynamically renders an electronic water level mimicking a physical spirit level based on pitch and roll calculations. In an emergency, it instantly transitions to a high-visibility flashing SOS banner.

Figure 6 — Interactive Displays: Electronic spirit-level water animation on Arduino UNO Q LED Matrix.

Battery Intelligence: The Hybrid SoC Algorithm

+─────────────────────────────────────────────────────────────────────────────+
| HYBRID SoC ALGORITHM LOGIC |
+─────────────────────────────────────────────────────────────────────────────+
| |
| 1. BOOT-UP: |
| • Read stabilized OCV from 14-bit ADC. |
| • Snap Coulomb counter (accumulatedAh) to OCV Lookup Table (LUT). |
| |
| 2. RUN-TIME DYNAMICS (Discharge / Charge): |
| • Coulomb Counting: ΔQ = I * Δt |
| • Charging Multiplier: 98% Coulombic efficiency |
| • Immunity to instantaneous voltage sags under motor PWM load |
| |
| 3. AUTOMATIC OCV RECALIBRATION: |
| • Rest Detection: Current |I| < 50 mA |
| • Relaxation Window: 60 continuous seconds at rest |
| • Action: Sync Coulomb Counter back to True Chemical OCV curve |
| • Zero Drift Accumulation over time! |
+─────────────────────────────────────────────────────────────────────────────+

Mathematical Formulation: During operation, charge is tracked second-by-second:

Capacity(t) = Capacity(t0) + ∫t0t η × I(τ) dτ

η = 1.00 during discharge (I < 0)
η = 0.98 during charge (I > 0) to account for Coulombic losses.
When the vehicle rests (|I| < 50 mA) for t_rest ≥ 60 s, the chemistry stabilizes, and the algorithm recalibrates the Coulomb counter against the non-linear Open-Circuit Voltage Lookup Table (OCV-LUT), eliminating sensor drift.

Deterministic RS-485 Protocol & 4-Step Fault Diagnostics

Custom Binary Packet Framing: Inter-node communication follows a deterministic master-slave polling cycle every 2.0 seconds:

+-----------+-----------+---------------+---------+--------------------+---------+-----------+
| SOF [0] | SOF [1] | Payload Len | Command | Payload | CRC | EOF |
| 0x55 | 0xAA | (1 Byte) | (1 Byte)| (N Bytes) | (1 Byte)| 0xAA |
+-----------+-----------+---------------+---------+--------------------+---------+-----------+

Autonomous 4-Step Fault Diagnostic State Machine: When communication ceases for > 30 seconds, EVA Guardian executes an internal diagnostic sequence to isolate the exact failure point:

Figure — Autonomous 4-Step Fault Diagnostic Routine Flowchart

Step 1: Master triggers local hardware loopback (CMD 0x02, payload 0x1010). If it fails
→ IDS RS-485 Transceiver Hardware Fault.

Step 2: If master loopback passes, it commands slave BMS to run its internal loopback. If slave fails
→ BMS RS-485 Transceiver Hardware Fault.

Step 3: If both local transceivers pass but inter-node communication fails
→ Communication Cable Fault / Disconnected Wire.

Power Harness Degradation & Automatic Power Failover

Contact Resistance Monitoring: Connectors develop contact resistance (R_contact) through oxidation and vibration. EVA Guardian performs continuous dual-point differential voltage sensing:

[ Battery Pack ] ──> (V_BMS = 8.20 V) ───[ Corroded Connector / Harness ]───> (V_IDS = 7.55 V)
 ΔV = V_BMS - V_IDS = 0.65 V → [ WebUI Warning: "Connector Degradation Detected!" ]

ΔV = V_BMS − V_IDS = 0.18 V → [WebUI Status: Healthy harness (Green status)]

ΔV = V_BMS − V_IDS = 1.18 V → [WebUI Error: Critical Power Loss: Connector degradation detected. Schedule maintenance before terminal overheating.]

Physical Scale Demonstrator & WebUI Dashboard

To validate EVA Guardian in a realistic physical environment, a scale-model electric vehicle chassis was constructed:

Spatial Separation: Rear compartment houses the primary 2S battery pack, HX-2S-A2 BMS, and Node 1 electronics.
Cockpit Integration: Front compartment houses the Incident Detection Node, backup battery pack, power transfer relay, and onboard matrix display.
Harness Routing: Distinct cable channels for high-current power lines and differential RS-485 twisted pairs.
Live WebUI Dashboard: Real-time web browser interface providing comprehensive telemetry: live pack voltage, current, temperature, SoC%, SoH%, 3D orientation visualizer, connector health status, and diagnostic error codes.

Assembled Scale Model EV Demonstrator & Hardware Gallery

Chassis & Harness Routing: Assembled scale EV demonstrator with the live system telemetry dashboard	Rear Powertrain & BMS Node: Integrated rear matrix display actively rendering telemetry data on the scale model.
Hardware Integration: Spatial separation of high-power traction rails and differential RS-485 communication bus.	BMS Module: 2S Li-ion pack, HX-2S protection board, and analog signal conditioning module.

Real-Time WebUI Telemetry & Diagnostics Dashboard

Battery Telemetry View: Real-time pack voltage, SOC%, SOH%, current, temperature, and connector health.

Incident Telemetry View: Continuous monitoring of abnormal motions, suggesting precautionary measures

Experimental Validation & Fault Injection Testing

The entire system was subjected to rigorous stress tests and deliberate fault injection:

Experiment	Injected Fault	Condition Observed	System Response	Validation Result
Dynamic Load Pulse	2.5 A discharge pulse via L298N driver	Voltage sagged by 0.5 V	Coulomb counter maintained steady 82% SoC without false drop	PASSED (Sag Immunity)
OCV Drift Correction	Simulated 15 mAh current integrator drift	After 60 s rest	Algorithm synced Coulomb count to chemical OCV curve	PASSED (Zero Drift)
Speed Bump / Pothole	High-g vertical transient shocks	PSD classifier correctly categorized as Up and Down	Spirit level remained active, 0 false alarms	PASSED (Zero False Alarm)
Crash & Tip-Over	Simulated collision followed by tilt	3-window voting confirmed Accident (94% conf)	Telegram alert sent in 1.2s; matrix flashed SOS	PASSED (Instant SOS)
Bus Cable Cut	Disconnected RS-485 wire	4-step diagnostic state machine triggered	Correctly isolated Cable Fault on dashboard	PASSED (Fault Isolation)
Connector Oxidation	Inserted series resistance in power harness	ΔV rose to 1000 mV	WebUI immediately flagged Connector Degradation	PASSED (Error Indication)
Traction Rail Collapse	Cut primary power supply during active trip	Relay switched to backup battery in <10ms	Safety node and Telegram uplink remained fully operational	PASSED (Fail-Safe Power)

Complete Video Demonstration Suite

The complete hardware validation and fault-injection scenarios have been recorded and documented in detail:

• Video 1: Comprehensive Technical Architecture & End-to-End Walkthrough

Description: Full walkthrough demonstrating dual-node RS-485 telemetry, Edge AI motion inference, live WebUI metrics, and emergency response workflows.

• Video 2: Fail-Safe Power Failover: Automatic Transfer Relay Under Main Rail Collapse

Description: Live demonstration showing the transfer to the backup 2S battery pack when the main traction rail is severed.

• Video 3: Autonomous 4-Step Communication Fault Diagnostics & Wire-Break Detection

Description: Deliberately cutting the differential bus cable and demonstrating how EVA Guardian isolates the fault to the harness rather than blaming transceivers.

• Video 4: Predictive Connector Wear: Real-Time Differential Voltage Sensing (ΔV)

Description: Introducing series contact resistance and observing real-time ΔV voltage drop warnings on the WebUI dashboard before thermal damage occurs.

• Video 5: Integrated Scale Electric Vehicle Demonstrator

Description: Final review of the integrated electronics housed within the scale electric vehicle model, showcasing harness segregation and cockpit indicators.

Bill of Materials (BOM)

#	Part Number	Manufacturer	Description	Qty	Reference / Link
1	ABX00162	Arduino	ARDUINO UNO Q 2GB RAM 16GB EMMC	2	element14 • UNO Q SBC
2	EVAL-ADM3068EEBZ	Analog Devices	EVAL BOARD FOR ADM3068E RS-485	2	element14 • ADM3068E Eval
3	MX150 Receptacle	Molex	Cable Assembly, MX150 Receptacle to Free End, 2 Ways, 600 mm	3	element14 • MX150 Cable
4	ACS723 Module	Smart Elex	SmartElex Current Sensor Breakout - ACS723 (Low Current ±5A)	1	Robu.in • ACS723 Breakout
5	L298N Motor Driver	Smart Elex	SmartElex 2x2A DC Dual Motor Controller V2 (L298N)	1	Robu.in • L298N Driver
6	18650 Battery Charger	Generic	2S 8.4V 2A 18650 Lithium Battery Charger Module Type C	1	Robu.in • 2S Type-C Charger
7	18650 Li-ion Cells	Generic	18650 Capacity Li-ion Battery Cells	4	—
8	Battery Holder	Generic	18650 x 2 Battery Holder with Cover and On/Off Switch	2	—
9	HX-2S-A2 BMS	Generic	HX-2S-A2 Circular 2S 8.4V BMS 18650 Protection Board	2	Robu.in • HX-2S-A2 Module
10	LF412	Texas Instruments	LF412 Dual JFET Input Operational Amplifier IC DIP-8	1	Robu.in • LF412 Op-Amp
11	5V SPDT Relay	Generic	5V SPDT Relay IC	1	—
12	5V Relay Module	Generic	5V 4-Channel Relay Module	2	Robu.in • 4-Ch Relay
13	GY-91 IMU	Generic	GY-91 MPU6500 + BMP280 7DOF IMU Sensor Module	1	Robu.in • GY-91 Sensor
14	NTC Thermistor	Generic	10k Ohm NTC Thermistor (Steinhart-Hart Calibrated)	1	Robu.in • 10k NTC

Lessons Learned, Challenges & Future Roadmap

Technical Challenges Overcome:

1. Managing Switching Noise in ADC Lines

Implementing unity-gain LF412 op-amp buffers and firmware-level Exponential Moving Average (EMA) filtering reduced measurement noise by over 90%.

2. False Positive Suppression in Crash Detection

Relying on single-frame IMU spikes caused potholes to mimic accidents. Transforming data into the frequency domain via Power Spectral Density (PSD) in Edge Impulse and requiring a 3-consecutive-window confirmation completely eliminated false triggers.

3. Thread Safety in RTOS

Concurrently updating sensor telemetry, running neural network inference, and refreshing the LED matrix initially caused display artifacts. Implementing Zephyr RTOS mutexes resolved all resource contention.

Conclusion & Acknowledgments

EVA Guardian demonstrates that smart, predictive, and crash-resilient safety features do not require expensive automotive supercomputers. By combining low-cost embedded hardware (Arduino UNO Q), thoughtful analog circuit design, robust differential communication (RS-485), and lightweight on-device Machine Learning (Edge Impulse), we can build electric vehicles that are genuinely self-aware.

From detecting microscopic contact resistance degradation to autonomously isolating severed cables and preserving emergency communication through a crash, EVA Guardian sets a new benchmark for light electric vehicle safety.

Special Thanks: A heartfelt thank you to the element14 Community, the EZ-EV Design Challenge organizers, and all fellow challengers for their invaluable feedback and support throughout this journey!